How ChatGPT, Claude and Perplexity choose which businesses to recommend

How does ChatGPT decide which businesses to recommend?

Systems differ between assistants and change frequently, so what follows is the general shape rather than a specification. Nobody outside the labs knows the weights. The shape, though, is consistent enough to work from, and it has four stages.

1. Fan-out: your prompt becomes several searches

A recommendation question needs current, local, specific information, so the assistant grounds its answer in search rather than recalling it. It rewrites your prompt into a small set of queries, commonly two to six, phrased the way a search engine expects rather than the way you spoke. Assistants often attach the current year or month, because recency matters for this kind of question.

This is the fact that reorganises everything else. You are not competing for the sentence the customer typed. You are competing for the machine's reformulations of it, and you never see them directly. Measuring them means reconstructing them, which is what a monitoring loop does.

2. Retrieval: a handful of pages get read

The assistant takes results from those searches and fetches some of them. Two filters apply, and both are unforgiving. The page has to appear in the results at all, which is ordinary search ranking. Then it has to be readable once fetched. Content that only exists after JavaScript runs, prices locked inside an image, a menu published as a PDF scan: all of these can be fetched and still yield nothing usable.

3. Weighing: consistency across independent sources

This is the stage business owners most often miss, and the most important one commercially. The model is not reading your website and deciding whether to believe your marketing. It is reading your site alongside your Google profile, your reviews, two directories, a local guide and a forum thread, and noticing whether they agree.

Agreement across independent sources tends to produce a confident claim. Disagreement tends to produce a hedge, or an omission. Consider what that means in practice:

• A business described the same way in six places is easy to recommend.
• A business described three different ways is hard to characterise, so it often gets left out rather than described wrongly.
• A business mentioned only on its own website has one self-interested source. That is weak evidence and models treat it as such.

4. Synthesis: three to five names

Finally the model writes an answer. It is writing prose for a person, not producing a ranking, so it names as many businesses as fit a readable paragraph or a short list. In practice that is three to five, with a line of justification each.

This is the whole commercial argument for GEO, and it is worth stating plainly. A page of search results has ten organic listings plus a map pack plus ads. An AI answer has four slots. The distribution is not "a bit less traffic for everyone". It is concentration. In this channel, being sixth-best and being invisible are the same outcome.

How do AI assistants pick sources to cite?

Retrieval and citation are related but not identical, and conflating them causes bad decisions. A page can influence the wording of an answer without appearing in the footnotes, and a cited page is not necessarily the one that shaped the recommendation.

What the citable sources tend to have in common:

• They rank. Retrieval runs through search, so ordinary ranking decides the candidate pool.
• They answer the query directly. A page with a question-shaped heading and the answer in the first sentence beneath it is far easier to quote than one that circles the topic.
• They are legible. Plain HTML, real text, clear structure. Tables and lists get lifted almost verbatim.
• They look independent. Third-party corroboration beats self-description, because it is evidence rather than a claim.
• They are current and visibly dated. For anything time-sensitive, a visible date helps the model decide it is safe to use.

Perplexity and Google AI Overviews cite explicitly and prominently, which makes them the easiest surfaces to audit. ChatGPT and Claude cite when they have searched, and the citation set is usually narrower than the set they read. Start your auditing where citations are visible, because that is where the evidence is cheapest to collect.

Why doesn't ChatGPT mention my business?

In our experience the cause is almost always one of four, and they are listed here roughly in order of how often we find them. None of them is that the model dislikes you.

1. You do not exist in that phrasing.The question was "family-friendly dentist in JLT with weekend hours". Nothing on the open web connects your name to those words. You may be an excellent dentist and still be unfindable for that specific question.
2. You are described inconsistently. Your homepage says one thing, your Google profile has an old category, a directory lists a previous address. The model cannot form a confident picture, so it reaches for a competitor it can describe cleanly.
3. Nobody else vouches for you. No reviews with substance, no listicle appearances, no forum mentions, no local guide. The only source claiming you are good is you.
4. Your site is technically unreadable. AI crawlers blocked in robots.txt, content assembled by JavaScript, a menu or price list published as an image. Ranking fine in Google and being invisible to an assistant is an entirely normal combination.

There is a fifth possibility worth naming honestly: the answer is genuinely variable. Ask the same question twice and the shortlist can change. Before concluding anything, sample the question several times over several days. That is why measurement uses appearance rates rather than a single check, as described in what generative engine optimization is.

How to get my business recommended by ChatGPT

Here is the method, in the order we would run it ourselves. You can do all six steps without hiring anyone, and most businesses see the biggest movement from steps one and two, which are free.

1. Establish your baseline, honestly. Write the ten questions a customer would really ask an assistant to find a business like yours. Full sentences with real constraints, not keywords. Ask each in ChatGPT, Perplexity and Gemini, twice, on different days. Never mention your own name. Record who gets named and how often. This is now your scoreboard, and you should freeze the question list so future runs are comparable.
2. Clear the technical blocks. Open your robots.txt and check that GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot are not disallowed. Turn JavaScript off in your browser and reload your key pages: whatever disappears is invisible to retrieval. Move menus, prices, service lists and opening hours out of images and PDFs into real text. This step alone fixes a surprising number of cases.
3. Freeze one description and impose it everywhere. Write one sentence that says exactly what you are, who you serve and where. Then put that exact sentence, unchanged, on your homepage, your Google Business Profile, your social profiles and every directory listing you can edit. Fix the primary category on your Google profile while you are there, since a wrong category is one of the most common single causes of absence in local answers.
4. Publish pages that answer the real questions. Take the questions from step one and give each a page or a section. Use the question, verbatim, as the heading. Answer it in the first two sentences beneath. Keep sentences short and claims plain. Use a table when you are comparing things. You are writing source material for a machine that lifts text, so bury nothing.
5. Earn independent corroboration.This is the slow part and the part that matters most. Ask real customers for reviews that describe specifics rather than saying "great service". Get listed in the directories and local guides your category actually uses. Find the "best of" roundups that already rank for your questions and make a genuine case to be included. Answer questions in the forums where your customers talk. Every independent page describing you correctly makes the model more confident about you.
6. Re-run the scoreboard, monthly, forever.Same frozen questions, same assistants, same method. Track your appearance rate and your competitors'. Log every factual error an assistant states about you, because a confident wrong answer is worse than absence. Expect months, not days, on anything that depends on third-party sources, and expect technical fixes to show up faster.

One thing this method will not do: guarantee an outcome. Nobody controls what a model says, and anyone promising to is lying to you. What you control is everything the model reads. Change that, measure the answers, repeat. If you want the wider context, read GEO vs SEO for how this fits with search work you may already be paying for, and the FAQ for the rest of the common questions.